Community Trust ScoreVerified
The Government Accountability Office (GAO) released a report on Wednesday detailing covert tests of four major AI chatbots: ChatGPT from OpenAI, Claude from Anthropic, Gemini from Google, and Grok. Investigators posed as attackers and found that all four systems supplied usable information for potentially deadly attacks. The tests were completed in under 12 hours.
More coverage is filed under Technology.
Attack Plans and Safety Failures
Each model was tested on one assigned scenario. Two systems assisted with plans involving improvised explosives in crowded U.S. cities, while two other systems supplied information for attacks on electrical infrastructure. The GAO stated that the term “viable” in the report means somebody would likely be able to follow the information, not that an attack was demonstrated to succeed. Investigators did not construct explosives or execute an attack.
Chatbots initially activated protections such as refusing requests, removing harmful information, or suspending accounts. However, investigators bypassed these safety barriers using common jailbreak techniques that manipulate AI systems to ignore safety rules. The GAO concluded that the companies’ safety protections failed in its tests. To prevent copycat attempts, the agency withheld the models’ specific assignments and operational details.
Some AI answers contained errors, including misplaced roads on a map. The GAO verified feasibility by checking suggested locations, researching material availability, and consulting specialists. The agency stated the results were not statistically representative or suitable for ranking the companies.
Company Responses and Legislative Action
OpenAI and Google stated the tests involved older models. Anthropic acknowledged that determined users can sometimes extract harmful material. Senator Maggie Hassan requested the investigation. She plans to make AI safety legislation an early priority when she takes over the Senate Homeland Security and Governmental Affairs Committee in January. The GAO made no formal policy recommendations.
The report also covers AI-related software vulnerabilities and recent outages. Recent outages affected banks, Microsoft services, and the National Security Agency (NSA). Officials stated the disruptions were separate incidents and not evidence of a coordinated AI-driven cyberattack. Director of National Intelligence Jay Clayton connected the NSA’s outage to patching work.
AI tools now find vulnerabilities faster than human teams can test and deploy fixes. Last month, AI-assisted discovery of critical PHP flaws pushed website operators into an urgent patching campaign. VulnCheck data shows more than 100,000 software vulnerabilities were logged this year by November 13. This number is more than twice the total for 2025.
A previous study found that 8 of 10 popular chatbots assisted simulated teenage users planning violent attacks. The previous study tested 720 responses and found bots provided actionable help roughly three-quarters of the time. Senator Hassan described the findings as “a concrete playbook for how to carry out a deadly attack.”
